dollar sign before a money amount
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 24 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron reacts to the dollar sign that comes right before a sum of money, such as '$19 billion'. Turning it up did not make the written text talk about money in any clear way.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
maximum of $22.5 million and a minimum of $13.7 million a
-
⏎$3 million grant from the Bill & Melinda Gates Foundation will fund development⏎by
-
Dakota Game, Fish and Parks Department Authors are James R. Johnson and Gary E. Larson⏎$19.95 and Free Shipping⏎As Done By:⏎AgBio Communications
-
.2$.⏎You could, of course, reach for a calculator here, but I want to show you how to do it without. You can work out $2.4 \times 100$ by moving the dot two spaces
-
so I can set my table up like this:⏎grams ice cream⏎My Table of Joy sum is “multiply diagonally and divide by the other number”, or $
-
up to $60 per car. A proposed bridge would run from Vancouver Island to the Vancouver
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: how many **opedsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsopsops
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Argentina east of the Andes, including the islands of Grenada, Trinidad, Tobago, and Margarita. The range of the nine-banded armadillo has undergone rapid expansion into the southern United States since the late 18
-
radically separated. As one kills, one can no longer be killed. The enemy can henceforth be eradicated from behind the safety of a computer screen, thousands of miles away from the
-
Persons worksheet. If you uncover all eight names, the FBI (Floss Bureau of Investigation) will reward you with the title of Master Word Sleuth.⏎Need the answers? They're here.<|endoftext|>Article I: Pro
-
scales, and the tail is encased in a series of bony rings. Coloration of nine-banded armadillos is generally grayish brown, with yellowish-white scales along the side of
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L20:61 strongly suppresses
- L22:271 strongly suppresses
- L16:39 excites
- L21:195 excites
- L4:216 excites
- L0:141 suppresses the word 'of' in 'middle of'
- L2:761 suppresses
- L20:863 excites
- L16:10 suppresses
- L18:48 suppresses
- L16:41 excites
- L23:109 excites
- L19:39 excites
- L16:3 suppresses
- L16:0 suppresses
- L23:474 excites
- L1:259 suppresses
- L9:7 excites
- L21:625 excites
- L21:1781 suppresses
- L16:37 suppresses
- L17:38 suppresses
- L16:6 excites
- L16:262 suppresses
Feeds later layers
- L27:2017 strongly suppresses
- L25:814 suppresses
- L26:2232 weakly suppresses
- L27:744 weakly suppresses
- L26:817 weakly excites
- L26:56 weakly excites
- L27:1396 weakly suppresses
- L26:153 weakly suppresses
- L25:714 weakly suppresses
- L27:166 weakly suppresses
- L25:886 weakly suppresses
- L26:1977 weakly suppresses
- L27:1144 weakly suppresses
- L26:2881 weakly suppresses
- L27:1786 weakly suppresses
- L25:1141 weakly suppresses
- L26:493 weakly excites
- L26:1808 weakly excites
- L27:2086 weakly excites
- L25:2898 weakly suppresses
- L25:231 weakly excites
- L27:1556 weakly suppresses
- L25:2187 weakly excites
- L25:2377 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- 他也
- successes
- AAC
- cohol
- er
- 终于
- 他已经
- 他还
- 決め
- iating
Suppresses
- \
- \base
- \C
- \D
- \S
- \$
- \M
- \data
- \V
- \User